A novel explainable deep learning framework for reconstructing South Asian palaeomonsoons

Abstract

Abstract. We present novel explainable deep learning techniques for reconstructing South Asian palaeomonsoon rainfall over the last 500 years, leveraging long instrumental precipitation records and palaeoenvironmental datasets from South and East Asia to build two types of model: dense neural networks ('timeline models') and convolutional neural networks (CNNs). The timeline models are trained individually on seven regional rainfall datasets and while they capture decadal-scale variability and significant droughts, they underestimate interannual variability. The CNNs, designed to account for spatial relationships in both predictor and target, demonstrate higher skill in reconstructing rainfall patterns and produce robust spatiotemporal reconstructions. The 19th and 20th centuries were characterised by marked inter-annual variability in the monsoon, but earlier periods were characterised by more decadal- to centennial-scale oscillations. Multidecadal droughts occurred in the mid-seventeenth and nineteenth centuries, while much of the eighteenth century (particularly the early part of the century) was characterised by above-average monsoon precipitation. Extreme droughts tend to be concentrated in south and west India and often coincide with recorded famines. By applying explainability techniques, we show that the models make use of both local hydroclimate and synoptic-scale dynamical relationships. Our findings offer insights into the historical variability of the Indian summer monsoon and highlight the potential of deep learning techniques in palaeoclimate reconstruction.